Your agent's memory is a dataset you can publish
funes indexes every Claude Code, Codex, pi, and Hermes session into one searchable memory, so your agent stops re-deriving decisions you already made.

What it does
funes turns past sessions from Claude Code, Codex, pi, and Hermes into a single local index — a Lance dataset built by one deterministic pipeline: parse each transcript into a shared turns format, chunk it, embed it with a pinned local model. Your agent gets recall and get as tools, and hooks keep the index current every turn, so mid-task it can pull up last week’s rationale without being asked. Every hit is labeled with the agent it came from, subagent transcripts included.
The interesting bit A memory here is a first-class Hub artifact: push it and the process behind a project — decisions, dead ends, rationale — becomes something a teammate, another machine, or anyone (if you make it public) can recall from. funes ships no generative model of its own; it does retrieval with a pinned local embedder and reranker, and you reason with whatever agent you already use. The name is a joke with a thesis: Borges’ Funes remembered everything and it wrecked him — this one chunks, reranks, and forgets differences on purpose.
Key highlights
- Cross-agent continuity: start a task in Claude Code, finish it in Codex, and each recalls the entire history — not just its own sessions.
- Hybrid retrieval: vector + BM25 fused, reranked, and reweighted by recency.
- Local inference: pinned embedding and reranking models on a built-in backend (Accelerate on macOS, pure Rust on Linux); the default build carries no ML runtime dependency, with an ONNX backend as an opt-in.
- Guarded publishing: credentials are redacted at index time, an always-on gate refuses to push any chunk still containing a secret, and Hub datasets are private by default.
askborrows a coding agent for one question and returns a grounded answer that names the sessions it drew from — nothing installed.
Caveats
- The embedding model is pinned and stamped into the memory; a different one is refused, and changing it means rebuilding from transcripts (raw text is retained in every row, so it’s a rebuild, not a loss).
- The automated flow — first index, per-turn hook, session-boundary publishing — is documented for Claude, Codex, and Hermes; pi gets the tools but less of the machinery.
- The secrets gate catches credentials, not judgment: a public memory shares your dead ends and half-formed rationale along with the good parts.
Verdict If you bounce between coding agents, or want your team’s reasoning history to be a queryable, shareable artifact, this is worth a close look. If you live in one agent and never revisit old sessions, it’s an elegant answer to a question you don’t ask.
Frequently asked
- What is huggingface/funes?
- funes indexes every Claude Code, Codex, pi, and Hermes session into one searchable memory, so your agent stops re-deriving decisions you already made.
- Is funes open source?
- Yes — huggingface/funes is open source, released under the Apache-2.0 license.
- What language is funes written in?
- huggingface/funes is primarily written in Rust.
- How popular is funes?
- huggingface/funes has 601 stars on GitHub.
- Where can I find funes?
- huggingface/funes is on GitHub at https://github.com/huggingface/funes.